HA causes reboot of xcp-ng nodes

john.c

Technical Summary: Weak Host Model & HA Heartbeat Failures

The Problem: Asymmetric Routing & The Weak Host Model

Linux defaults to a weak host model for IPv4 networking. The kernel treats IP addresses as belonging to the entire host, rather than being strictly bound to a specific physical or logical network interface (NIC).

When a multi-homed cluster node transmits High Availability (HA) heartbeats or storage synchronization packets, the kernel evaluates its local routing table. If multiple paths exist to the destination subnet, the kernel will choose the egress interface based on the lowest routing metric (cost).

This causes two major issues in HA clusters:

Asymmetric Outbound Traffic: The kernel may send HA packets out of a secondary interface (e.g., a storage or management NIC) instead of the dedicated HA/heartbeat interface.
ARP Flapping & Confusion: Because of the weak host model, a node might reply to ARP requests for its HA IP address on any active interface. This poisons the ARP tables of adjacent switches and cluster peers, misrouting traffic to the wrong physical ports.

Expected Path:  [Node A: Dedicated HA NIC] ---------> [Switch] ---------> [Node B: Dedicated HA NIC]
Actual Path:    [Node A: Management NIC]   -- (Lowest Metric Route) --> [Node B: Management NIC]

The Impact on HA Clusters (XCP-ng / XAPI)

When heartbeats leak onto the wrong network or get dropped due to strict firewalling (e.g., a management network blocking cluster traffic), the cluster experiences a false split-brain condition.

Peer nodes assume the host is dead because heartbeats stopped arriving on the designated HA network.
The HA fencing mechanism is triggered, causing the host to spontaneously reboot or self-fence to protect shared storage from corruption.

Recommended Subsystem Remedies

To enforce a strong host model where traffic strictly respects interface boundaries, the kernel and network orchestration layer (XAPI) must implement specific controls:

Strict ARP Filtering (sysctl)
Prevent the kernel from answering ARP requests for an IP on the wrong interface.
```
net.ipv4.conf.all.arp_ignore = 1
net.ipv4.conf.all.arp_announce = 2
```
Reverse Path Filtering (RP-Filter):
Ensure the kernel drops packets that arrive on an interface if the return path would not normally use that same interface.
```
net.ipv4.conf.all.rp_filter = 1
```
Policy-Based Routing (PBR):
Configure separate routing tables for the HA interface so that traffic sourced from the HA IP is explicitly forced out of the HA NIC, ignoring global metrics in the main routing table.

So outside of altering these settings, if not altered at kernel source code level with a patch (that’s upstreamed to Xen Project), it will default to weak host model, a patch is required for it to instead default to the strong host model. If necessary it can target management, storage, backup, migration and/or HA networks specifically.

Doing a kernel source code patch will especially help out TWINSTOR development, currently in technical preview!

john.c

@tjkreidl When you’re back online my posts above, may reveal something about behaviours you noticed in Citrix XenServer and Citrix Hypervisor, during your employment there as CTP.

@poddingue This is a potential source for a joint patch to Xen Hypervisor kernel to fix this as part of upstreaming, something that will benefit from multiple eyes and hands working on it.

tjkreidl

@john.c Interesting information about the strong v. weak host model.
Active-active and LACP bonds will split the network traffic between the NICs, while an active-passive bond only uses the primary. WHat type of bonds do you have set up?

john.c

@tjkreidl said:

@john.c Interesting information about the strong v. weak host model.
Active-active and LACP bonds will split the network traffic between the NICs, while an active-passive bond only uses the primary. WHat type of bonds do you have set up?

@tjkreidl It is either a balance-alb (Mode 6) or a balance-xor (Mode 2) bond topology on this specific deployment (as the upstream router has limited port capacity and does not support LACP).

However, the brilliant part about this architectural issue is that the underlying Layer 3 routing vulnerability remains identical regardless of which of these two modes is active. Both modes open up multiple physical paths for outbound traffic under a single logical bond, giving the kernel's default weak host model the perfect opportunity to misroute packets under load.

Here is how the weak host model breaks both configurations:

If it is balance-alb (Mode 6): The bonding driver actively performs Layer 2 ARP and MAC manipulation to balance paths without switch assistance. The weak host model completely undermines this logic because the kernel treats the IP address as globally accessible to the whole host. Under load, the kernel's Layer 3 routing engine completely ignores the bonding driver's intended pathing boundaries, picks a "cheaper" path via global metrics, and leaks the packet out of a completely separate infrastructure interface (like Management or Storage).
If it is balance-xor (Mode 2): The driver relies on a strict hash policy (like layer2 or layer2+3) to statically map traffic to a destination across a specific physical NIC slave. Yet, if a heavy background process (like a backup job or storage replication) alters local routing table costs or causes transient congestion, the kernel’s Layer 3 logic overrides that static Layer 2 pathing—spilling packets out of an unrelated physical port.

Ultimately, OpenMetrics packet-flow telemetry caught this exact moment of divergence: the Layer 3 stack bypassed the logical bond boundary entirely. The packet exited on an unintended infrastructure port carrying the wrong source IP, where adjacent switches or firewalls dropped it as unroutable. To XAPI and the HA daemon, the heartbeat was instantly lost on that specific port, triggering the self-fencing reboot loop.

This is why it's a universal vulnerability across multi-path bonding modes on a multi-homed system, and why an upstream patch enforcing a strong host model for the Infrastructure Plane is the cleanest solution.

john.c

@AtaxyaNetwork Depending on your current or past homelab topography, this behavior might ring a few bells for you as well.

Homelab and prosumer environments are highly susceptible to this exact weak host routing leak. Because we often multiplex distinct infrastructure planes (Management, dedicated storage networks, and backup backplanes) across multi-port NIC bonds—frequently using balance-alb or balance-xor because the upstream switches lack stacked enterprise LACP support—the conditions are perfect for a routing collision.

If a heavy data operation (like a massive VM migration or a backup sync) alters the local metric weightings or causes micro-congestion, the default weak host model can silently push Dom0 host-terminated packets onto the wrong physical interface segment. It's a classic hidden variable that can cause erratic connection drops or unexplainable HA timeouts on otherwise perfectly configured hardware.

tjkreidl

@john.c Most interesting, and I fully agree, that a strong host model enforcement policy is really the best and only recourse for such a topology, or so it would seem.
The only other option that comes to mind would be to not put the heartbeat connection on any sort of bond or multipath. That's, of course, not ideal.

john.c

@tjkreidl said:

@john.c Most interesting, and I fully agree, that a strong host model enforcement policy is really the best and only recourse for such a topology, or so it would seem.
The only other option that comes to mind would be to not put the heartbeat connection on any sort of bond or multipath. That's, of course, not ideal.

@tjkreidl Exactly, and you have hit on the exact architectural compromise we've all been forced to make for years.

Moving the heartbeat connection off a bond/multipath and onto a dedicated, single physical NIC does reduce the Layer 3 path alternatives that trigger the weak host leak. However, as you rightly pointed out, it's highly non-ideal. By removing the bond, we introduce a single point of failure (SFP/cable/switch port) directly into the critical HA backplane. We shouldn't have to sacrifice physical hardware redundancy just to keep the kernel's Layer 3 routing engine from misbehaving.

This is precisely why enforcing a strong host model for the Infrastructure Plane is the true architectural solution. It allows administrators to safely use balance-alb, balance-xor, or multipathing for maximum hardware resilience, while ensuring that Dom0-terminated traffic strictly honors its designated interface boundaries regardless of global metrics.

Since we are in full agreement on the root cause and the ideal fix, this looks like a prime candidate for a strategic architectural shift. Hopefully, @TeddyAstie and the @Team-Hypervisor-Kernel can look at how we can implement this structural protection by default—perhaps utilizing targeted sysctl overrides (arp_ignore, arp_announce, rp_filter) or interface socket-binding for host-terminated infrastructure networks.

carloum70

@john.c now I remember, during the installation we used the eth4 as the management interface.

[16:54 dacshyp002 ~]# cat /etc/firstboot.d/data/management.conf 
LABEL='eth4'
MODE='static'
IP='172.28.4.11'
NETMASK='255.255.252.0'
GATEWAY='172.28.4.1'
MODEV6='none'
DNS='8.8.8.8'

Afterwords we created a bond bond0 with eth2/eth3 and used this as the management interface.
The xe pif-list command shows the following:

device ( RO)                   : bond0
               management ( RO): true
    IP-configuration-mode ( RO): Static
                       IP ( RO): 172.28.4.11

device ( RO)                   : eth4
               management ( RO): false
    IP-configuration-mode ( RO): Static
                       IP ( RO): 172.28.4.11

Is it that easy to remove the IP address from the eth4 interface? Because it make no sense to have duplicate ip-addresses.
I think we've had duplicate IP addresses all along, but we only noticed the issue after enabling HA.
By the way what is the correct way to change the management interface? We used the xsconsole.

john.c

@carloum70 said:

@john.c now I remember, during the installation we used the eth4 as the management interface.
[16:54 dacshyp002 ~]# cat /etc/firstboot.d/data/management.conf 
LABEL='eth4'
MODE='static'
IP='172.28.4.11'
NETMASK='255.255.252.0'
GATEWAY='172.28.4.1'
MODEV6='none'
DNS='8.8.8.8'
Afterwords we created a bond bond0 with eth2/eth3 and used this as the management interface.
The xe pif-list command shows the following:
device ( RO)                   : bond0
               management ( RO): true
    IP-configuration-mode ( RO): Static
                       IP ( RO): 172.28.4.11

device ( RO)                   : eth4
               management ( RO): false
    IP-configuration-mode ( RO): Static
                       IP ( RO): 172.28.4.11
Is it that easy to remove the IP address from the eth4 interface? Because it make no sense to have duplicate ip-addresses.
I think we've had duplicate IP addresses all along, but we only noticed the issue after enabling HA.
By the way what is the correct way to change the management interface? We used the xsconsole.

If you’re using a bonded NIC what mode is used? If not bonding it will allow for multiple NICs, via bond0-12 etc to show a single NIC. But only use once my above fix has landed, to avoid a return to this issue, unless forcing strong host yourself before hand.

Remove the IPs with the following process:-

xe pif-list params=uuid,device,IP,management
xe pif-reconfigure-ip uuid=<PIF_UUID> mode=none
xe-toolstack-restart

carloum70

@john.c We are using a LCAP bond.
For my understanding, if I remove the IP settings from eth4, the problem will still not be solved because of the known issue with the bond that you explained in “Weak Host Model & HA Heartbeat Failures.”
(At the moment we have HA disabled.)

john.c

@carloum70 said:

@john.c We are using a LCAP bond.
For my understanding, if I remove the IP settings from eth4, the problem will still not be solved because of the known issue with the bond that you explained in “Weak Host Model & HA Heartbeat Failures.”
(At the moment we have HA disabled.)

What use is eth4 put to is it still management, if so put it into bond0? In one of my earlier posts on this I also detailed how to in the meantime use sysctl to force strong host, but best to wait for the official update from Vates.

bvitnik

@carloum70 I can't really help you with anything specific but I must make a note that when analyzing network setup you must also take into account the Open vSwitch. Looking at the classic Linux networking, bridging and routing facilities is not enough to get you the full picture.

For example, if I remember correctly, Active-Active bonding (non LACP) is implemented by Open vSwitch with so called balance-slb mode. This mode is not available with classic Linux kernel bonding options. What you are seeing (duplicated IPs) is maybe perfectly normal for this kind of setup with Open vSwitch.

I would search for the cause of the issue somewhere else like network congestion. Heart beats are very sensitive to network congestion and can easily fail if you do not dedicate network links for this kind of traffic. If you are sharing the links over which heart beats are sent with some high traffic stuff (storage maybe?), it can easily make problems.

UPDATE: unfortunately I don't have access to these kind of bonded setups any more so I can't check if the configuration differs in any way from yours. My home lab thingies are single network interface only.

carloum70

@bvitnik I think there is some confusing.
During the initial installation we have have used the eth4 (172.28.4.11) as the management interface. Later on we created a bond with eth2 and eth3 and changed the management interface to this bond using the xsconsole, and also used the same ip-address (172.28.4.11). We thought that the IP address had also been removed from eth4 during the reconfiguration, but we didn't check it afterwards. It turns out that this was not the case.
So now I removed the ip from eth4 using the xe pif-reconfigure-ip command.

Next step is to use the sysctl workaround @john.c mentioned in an earlier post and enable HA again.

(I also think it would be wise to use dedicated bond for the heartbeat.)

john.c

@bvitnik said:

@carloum70 I can't really help you with anything specific but I must make a note that when analyzing network setup you must also take into account the Open vSwitch. Looking at the classic Linux networking, bridging and routing facilities is not enough to get you the full picture.

For example, if I remember correctly, Active-Active bonding (non LACP) is implemented by Open vSwitch with so called balance-slb mode. This mode is not available with classic Linux kernel bonding options. What you are seeing (duplicated IPs) is maybe perfectly normal for this kind of setup with Open vSwitch.

I would search for the cause of the issue somewhere else like network congestion. Heart beats are very sensitive to network congestion and can easily fail if you do not dedicate network links for this kind of traffic. If you are sharing the links over which heart beats are sent with some high traffic stuff (storage maybe?), it can easily make problems.

UPDATE: unfortunately I don't have access to these kind of bonded setups any more so I can't check if the configuration differs in any way from yours. My home lab thingies are single network interface only.

The Open vSwitch would have been already taken into account with packet data flow analysis. Though if balance-slb is Open vSwitch exclusive analysis will be needed, but if influenced by kernel then it will also likely be affected.

stormi

CCing @Team-XAPI-Network and @Team-Hypervisor-Kernel for them to have a look at this strong vs weak host model and the impact on HA.

carloum70

Some update:

After removing the ip-address from eth4 the following messages disappeared

 [warn] SC: (script_service_do_query_liveset) reporting "heartbeat approaching timeout". host[1].time_since_last_hb=30408.
[info] SC: (script_service_do_query_liveset) "Heartbeat approaching timeout" turned FALSE

By the way I forgot to mention that each time I enable HA the following warning appears in xha.log

Sep 29 13:31:56 CEST 2026 [warn] BM: cannot open bonding status file (/proc/net/bonding/bond0). (2)
Sep 29 13:31:56 CEST 2026 [info] BM: this is not bonded. Terminating bonding thread.

bond0 (eth2 + eth3) is the management interface and also used for the HA.

tjkreidl

@carloum70 One option to address the HA error messages: Delete the HA configuration and re-configure it from scratch once the bonds are all configured OK, which it appears you think they now are.

john.c

@carloum70 said:

Some update:

After removing the ip-address from eth4 the following messages disappeared
 [warn] SC: (script_service_do_query_liveset) reporting "heartbeat approaching timeout". host[1].time_since_last_hb=30408.
[info] SC: (script_service_do_query_liveset) "Heartbeat approaching timeout" turned FALSE 
By the way I forgot to mention that each time I enable HA the following warning appears in xha.log
Sep 29 13:31:56 CEST 2026 [warn] BM: cannot open bonding status file (/proc/net/bonding/bond0). (2)
Sep 29 13:31:56 CEST 2026 [info] BM: this is not bonded. Terminating bonding thread.
bond0 (eth2 + eth3) is the management interface and also used for the HA.

Warning: Unconfigured Backup/Migration Networks Trigger HA Reboots

Watch out if a dedicated backup or migration network is not configured in Vates VMS, all of that traffic defaults to the management network.

When massive backup jobs or VM migrations saturate the management interface, it chokes the High Availability (HA) cluster heartbeats. This network congestion causes packet loss, triggers false split-brain conditions, and forces the host into a spontaneous reboot (self-fencing). This will especially occur on highly busy and congested instances of Vates VMS.

While sysctl ARP filtering patches the routing leak, it cannot fix physical network saturation.

The Long-Term Fix: Quad-Port Ethernet Upgrades

I highly recommend upgrading all server LAN cards (especially XCP-ng hosts) to multiple quad-port modules across multiple slots or daughter cards. Abundant physical interfaces allow you to build dedicated LACP or Active-Backup bonds, ensuring complete infrastructure isolation:

• Dedicated HA & Management: Keeps critical XAPI orchestration and heartbeat checks isolated from heavy data bursts.
• Dedicated Storage: Keeps iSCSI, NFS, or XOSTOR disk I/O running smoothly on its own low-latency pipe.
• Dedicated VM Traffic: Isolates production guest network traffic from infrastructure management tasks.
• Dedicated Backup & Migration: Physical ports can be carved out exclusively for VM motions and backup windows—or at least provide a dedicated, shared channel away from HA traffic.

tjkreidl

@john.c Keeping the various network traffic isolated according to specific usage (management, storage, VMs, etc.) is always a good idea. The last system I managed had 10GiB LACP bonds and using VLANs to isolate traffic and that worked fine with a four-node pool running around 80 or so XenDesktop instances per node. Never experienced any congestion issues. Each dom0 had a ton of memory and I believe it was either 8 or 16 VCPUs to make sure there were sufficient compute and memory allocations to allow for sometimes very heavy loads. It also helped that I eventually added GPUs to take on some of the computational load, in particular when some of the VMs were running applications employing heavy graphics.

john.c

@tjkreidl said:

@john.c Keeping the various network traffic isolated according to specific usage (management, storage, VMs, etc.) is always a good idea. The last system I managed had 10GiB LACP bonds and using VLANs to isolate traffic and that worked fine with a four-node pool running around 80 or so XenDesktop instances per node. Never experienced any congestion issues. Each dom0 had a ton of memory and I believe it was either 8 or 16 VCPUs to make sure there were sufficient compute and memory allocations to allow for sometimes very heavy loads. It also helped that I eventually added GPUs to take on some of the computational load, in particular when some of the VMs were running applications employing heavy graphics.

@tjkreidl Thanks Tobias, that’s great validation! Running 10GiB LACP bonds with proper VLAN isolation is definitely the ultimate goal for production stability, especially when pushing 80+ VMs per node.

Your point about dom0 resource allocation is also huge—people often forget that saturated vCPUs and starved dom0 memory can bottleneck network processing just as fast as a saturated physical link under heavy loads. Giving dom0 the extra compute headroom ensures the orchestration layer doesn't drop packets when backups or migrations scale up across that many instances.